def _store_entry(self, index, linear, w):
hp = self.hparams
basename = self._waves[index].with_suffix(".npy").name
seq_len = hp.mel_len * hp.r
lin_len = linear.shape[1]
assert lin_len <= seq_len # sanitary check
# mel spectrograms
mel = np.dot(self.mel_basis, linear) # transform to mel scale
mel = self._padding(mel.T, (seq_len, hp.n_mels))
mel = np.log(np.clip(mel, 1e-5, None)) # normalize to log scale
mel = mel.reshape(-1, hp.n_mels * hp.r)
np.save(self._path / 'mel' / basename, mel)
# sequence of character
text = text_normalize(self._texts[index], self.hparams.vocab)
assert len(text) < hp.text_len
text = text.ljust(hp.text_len, '~')
text = [self._char2idx[ch] for ch in text]
np.save(self._path / 'text' / basename, text)
# gate
gate = np.zeros(hp.mel_len, dtype=np.int)
gate[lin_len // hp.r:] = 1
np.save(self._path / 'gate' / basename, gate)
def _store_entry(self, index, linear, w):
hp = self.hparams
basename = Path(self._waves[index]).with_suffix(".npy").name
seq_len = hp.n_frames * hp.r
lin_len = linear.shape[1]
assert lin_len <= seq_len # sanitary check
# compute linear spectrogram
post_linear = self._padding(linear.T, (seq_len, linear.shape[0]))
post_linear = audio.normalize(post_linear, hp)
np.save(self._path / 'linear' / basename, post_linear)
# compute mel spectrogram
mel = np.dot(self.mel_basis, linear) # transform to mel scales
mel = self._padding(mel.T, (seq_len, hp.n_mels))
mel = audio.normalize(mel, hp)
mel = mel.reshape(-1, hp.n_mels * hp.r)
np.save(self._path / 'mel' / basename, mel)
# compute text sequence
text = text_normalize(self._texts[index], self.hparams.vocab)
assert len(text) < hp.text_len
text = text.ljust(hp.text_len, '~')
text = [self._char2idx[ch] for ch in text]
np.save(self._path / 'text' / basename, text)
def synthesize(args):
char2idx = {ch: i for i, ch in enumerate(hp.vocab)}
with open(args.f_text, 'r') as file:
text = ''.join(file.readlines())
# normalize the text
text = text_normalize(text, hp.vocab)
if len(text) >= hp.text_len - 1:
text = text[:hp.text_len - 1]
text += '~' * (hp.text_len - len(text))
text = np.array([char2idx[ch] for ch in text]).reshape(-1)
hp.batch_size = 1
# load the model
model = Tacotron(hp)
model.training = False
model.load_parameters(args.f_model)
x_txt = nn.Variable([hp.batch_size, hp.text_len])
_, mag, _ = model(x_txt)
x_txt.d = text[np.newaxis, :]
mag.forward(clear_buffer=True)
wave = synthesize_from_spec(mag.d[0].copy(), hp)
wavfile.write(args.f_output, rate=hp.sr, data=wave) # write a sample
def synthesize(args):
char2idx = {ch: i for i, ch in enumerate(hp.vocab)}
with open(args.f_text, 'r') as file:
text = ''.join(file.readlines())
# normalize the text
text = text_normalize(text, hp.vocab)
if len(text) >= hp.text_len - 1:
text = text[:hp.text_len - 1]
text += '~' * (hp.text_len - len(text))
text = np.array([char2idx[ch] for ch in text]).reshape(-1)
hp.batch_size = 1
# load the model
model = Tacotron2(hp)
model.training = False
model.load_parameters(args.f_model)
x_txt = nn.Variable([hp.batch_size, hp.text_len])
_, mels, _, _ = model(x_txt)
x_txt.d = text[np.newaxis, :]
mels.forward(clear_buffer=True)
m = mels.d.copy().reshape(1, -1, hp.n_mels)
np.save(args.f_output, m.transpose((0, 2, 1)))
